prefer-direct-imports-over-module-mocks

Prefer direct unit imports over module mocks in tests

When you reach for vi.mock / vi.hoisted / vi.resetModules / vi.setConfig to make a piece of production code testable, stop and ask: can the production code be factored so the test imports the unit it cares about directly?

The usual symptoms that you're about to write the wrong shape:

• A test file with vi.mock('../../src/some-helper') because the unit under test calls some-helper as a side effect.
• A test file with vi.resetModules() in beforeEach to force re-evaluation of a module whose state leaks across tests.
• A test file with vi.setConfig({ testTimeout, hookTimeout }) at module scope because cold-importing the unit on every test is too slow.
• A test file with await import('…') inside a helper "so the mock is wired before module evaluation" — vi.mock is hoisted by vitest; a top-level static import sees the mock just fine. This pattern is almost always a misconception.

These are all signals that the production code has an entanglement (typically: config loading, environment access, time, or a stateful singleton) that's making the testable core hard to reach.

The fix

Extract a pure-ish core from the entangled wrapper:

• Pure-ish core function — takes the resolved inputs directly (e.g. { migrationsDir, spaceFilter }, not await loadConfig()), returns a result object or throws structured errors. No side effects beyond what the test's tmpdir or fixture owns.
• Thin wrapper — the CLI command builder / HTTP handler / scheduler hook — calls the entanglement (loadConfig, process.env, Date.now(), the singleton getter), resolves inputs, delegates to the core. One screen of code, no branching.

Tests import the core directly. Zero vi.* plumbing. If the wrapper is genuinely just resolve-inputs-and-call-core, you can skip wrapper tests entirely — the build closes the gap.

When module mocks are legitimately the right tool

Some entanglements can't be cleanly extracted: